Unknown vector controls touching at least one declared enclosure coordinate.
Millimetre registry in input order, preserving all fixed entries exactly.
Accepted data QR diagonal ratio; neither this nor a Gram pivot is a condition number.
Column-scaled, twice-reorthogonalized data-row rank, or null at the work limit.
Largest dependent-control excursion outside the original millimetre enclosure.
Unobserved controls and incomplete convergence remain explicit; rank is not certified.
Maximum over rows supported entirely by fixed coefficients, millimetres.
Maximum over rows with positive unknown coefficient support, millimetres.
Smallest accepted over largest Cholesky pivot; not a singular-value condition number.
Actual stored sparse Hessian entries, including symmetric partners.
Actual projected-gradient iterations.
Generated patch and local parameters of the largest checked correspondence.
True when every coefficient contributing to this correspondence is fixed by external authority.
Patch identity or zero-based patch ordinal, according to this record, addressing the supplied coefficient-patch registry.
Normalized u coordinate in the declared local patch or unit source chart.
Normalized v coordinate in the declared local patch or unit source chart.
Largest supplied point-correspondence residual, millimetres.
Largest assembled reduced observation stencil; the explicit row-work cap includes exact-zero terms.
One private worst correspondence and its original affine feasible range; never consumer geometry.
Fixed constants and independent-control terms expressing each retained original control.
Signed Bernstein row weights in original control order; derivative rows are not position convex combinations.
Fixed contribution to the affine coefficient row, in registered millimetres.
Indices into the shared control registry; bicubic nets use sixteen IDs with u varying fastest.
Bernstein control positions in the common registered millimetre frame, ordered by control registry ID.
Evaluated fitted position at the retained correspondence, in registered millimetres.
Global normalized chart coordinates, or null when the patch has no declared global bounds.
Global normalized bounds of the containing refinement leaf, or null when no leaf layout is supplied.
Normalized local patch coordinates [u,v] of the retained correspondence.
Lower physical support or componentwise enclosure bound, in the declared registered millimetre frame.
Always null: this coefficient-fit witness does not claim original source-face identity.
Patch identity or zero-based patch ordinal, according to this record, addressing the supplied coefficient-patch registry.
Declared patch identity, or null when the caller supplied only positional patch indices.
Single-row affine-box feasibility observation; it does not certify joint fit or continuous source support.
Binary64 sum of absolute combined weights, in first-address order.
Predicate scope and diagnostic precision qualification.
Exact strict distance-squared comparison for the combined binary64 row.
Outward-rounded lower endpoints, millimetres.
Approximate distance to the exact row image, millimetres.
Unique original addresses, including exact-zero cancellations.
Outward-rounded upper endpoints, millimetres.
Binary64 sum of combined weights, in first-address order.
Original position/physical-gradient correspondences used to diagnose this patch's retained worst witness.
Observed target in the operation's millimetre exchange scale; derivative rows retain their physical directional meaning.
Sparse affine contributions in original control IDs; duplicate IDs and signed weights retain the owning operation's declared combination rule.
Upper physical support or componentwise enclosure bound, in the declared registered millimetre frame.
Position-plus-physical-gradient objective RMS in the declared millimetre exchange scale.
Numerical Gram rank, or null when the explicit dense diagnostic limit is exceeded.
Binary64 tangent-plane box gap for the assembled normalized objective; not a rigorous error-bounded certificate.
Tangent-plane support gap over the feasible normalized box; a numerical bound, not an outward-rounded certificate.
Infinity norm of the original normalized objective gradient.
Constant term of the assembled normalized least-squares objective.
Physical millimetre origin subtracted before the normalized solve.
Positive millimetre scale dividing physical position residuals in the normalized solve.
Current normalized quadratic objective value under the declared row weights.
Reported objective-gap upper estimate restored to square-millimetre scale.
Per-patch maximum restricted to rows with unknown coefficient support.
Actual local UV of each patch maximum, used by the declared residual refinement wave.
Per-patch maximum over fit and additional validation correspondences, millimetres.
Millimetre exchange-length-scaled physical-gradient RMS, separate from position RMS.
Actual transient signed derivative rows assembled without expanding the observation table.
Infinity norm of the normalized projected-gradient mapping.
True only when numerical full observation rank, gradient termination and supplied residual pass.
Relative Gram pivot threshold; this is not an exact rank certificate.
Affine substitution accounting; this does not certify convex weights.
Deepest affine dependency expansion actually traversed.
Largest retained affine substitution expression in original control IDs.
Largest absolute deviation of an affine expression's weight sum from one.
Smallest retained affine weight; negative values remain explicit and do not imply convexity.
Weighted RMS of all supplied point correspondences, millimetres.
Original and diagonally transformed spectral estimates and actual restart count, not condition numbers.
Largest Hessian diagonal used in scaling and diagnostics.
Smallest positive Hessian diagonal; null means no positive entry was observed.
Row-sum spectral upper estimate for the original Hessian, not a condition number.
Actual acceleration-momentum resets after objective increase.
Row-sum spectral upper estimate after diagonal congruence, not a condition number.
Unknown coefficient count after fixed shared controls are eliminated.
Controls with no positive observation basis support, independent of rank work limits.
Finite fitted coefficients with independent convergence and observation accounting.